GGSS: Geodesic-Gated Spherical Steering for Inference-Time Debiasing of Generative Vision-Language Models
GGSS cuts measured demographic bias across four tested generative VLMs without materially reducing general VLM accuracy.
The paper proposes an inference-time debiasing method that steers visual tokens on a unit hypersphere, using a gate to focus on tokens with stronger demographic signal. It tests four generative vision-language models against ten adapted debiasing baselines and prompt-based mitigation. GGSS reports the lowest average bias on all four models, with statistically significant gains on three backbones. MMStar accuracy stayed within +/- 0.6 percentage points of the unsteered baseline. HF Daily Papers' note
The paper proposes an inference-time debiasing method that steers visual tokens on a unit hypersphere, using a gate to focus on tokens with stronger demographic signal. It tests four generative vision-language models against ten adapted debiasing baselines and prompt-based mitigation. GGSS reports the lowest average bias on all four models, with statistically significant gains on three backbones. MMStar accuracy stayed within +/- 0.6 percentage points of the unsteered baseline. HF Daily Papers' note
score 4